As the AI wave continues to drive innovation in data center architecture, Broadcom recently announced the launch of the third-generation co-packaged optics (CPO) Tomahawk 6 "Davisson" network switch and the industry's first 800G Ethernet card, the "Thor Ultra," providing higher-performance networking solutions for next-generation AI computing clusters. Manish Mehta, Vice President of Marketing and Operations for Broadcom's Optical Systems Division, and Hasan Siraj, Head of Software and Ecosystem, emphasized the importance of promoting AI infrastructure development through an "open" approach during a discussion with Taiwanese media.
Third Generation CPO Technology: Tomahawk 6 Davisson
Manish Mehta, vice president of Broadcom's optical systems division, said in a briefing that the Tomahawk 6 Davisson is the industry's first 100T CPO Ethernet switch, with features including a transmission rate of up to 200G/lane, a total bandwidth of 102.4T, and a 70% reduction in power consumption compared to pluggable optical modules.
At the same time, the Tomahawk 6 Davisson is a reliable platform based on the verification of previous generation products. It has achieved zero link jitter in Meta's actual testing of one million device operation hours, emphasizing the stability of data exchange during its network transmission operation.
Manish Mehta emphasized that the reliability of CPO technology has been proven, and Meta's public test data shows that CPO technology improves maintenance efficiency by 5 times compared to pluggable optical modules, and can even increase training efficiency by 90% in a 24K GPU cluster.
Revolutionary Network Card: Thor Ultra
Hasan Siraj, head of software products in Broadcom's core switching division, introduced that Thor Ultra is an 800G Ethernet network card designed specifically for AI back-end training. Not only is it the industry's first single network card capable of 800G throughput, its overall operating power consumption is controlled at just 50W. It is also fully compliant with the Ultra Ethernet Consortium (UEC) 1.0 specification, supports packet-level multipath transmission and out-of-order packet processing, and offers both OCP 3.0 and PCIe specifications. It is also compatible with the next-generation PCIe Gen 6 design, meaning it can connect to future network transmission architecture deployment applications.
Hasan Siraj further pointed out that traditional RDMA protocols face challenges when used with large clusters of hundreds of thousands of XPUs. Thor Ultra, however, uses modern technology to address key issues such as multipath transmission and out-of-order processing. This improves the network latency and slowdown that can occur when large amounts of data, such as AI, are transmitted through servers for computation.
Complete AI Network Solution
The overall solution demonstrated by Broadcom shows that the combination of Tomahawk 6 switches and Thor Ultra network cards can support large clusters of up to 128000 GPUs in a two-layer network architecture, reducing power consumption, latency and optical component usage by more than 40% compared to traditional three-layer architecture.
Technology Roadmap and Industry Cooperation
During the Q&A session, Broadcom executives not only explained that the fourth-generation 400G/lane CPO product is already in the development stage, but also expressed their close cooperation with TSMC on the COUPE silicon photonics platform process. They also expressed their in-depth cooperation with Taiwanese companies such as Hon Teng Precision Industry Co., Ltd. and Delta Electronics in the fields of connectors, chassis and advanced packaging.
Furthermore, Intel recently demonstrated at the OCP Summit that it integrated its Xeon 6 server processors, Gaudi 3 accelerators, and NVIDIA's B200 GPU in the form of NVIDIA NVLink Fusion, using Broadcom's Tomahawk 5 51.2Tb/s network switch chip. Broadcom is optimistic about this configuration and believes that as the market becomes more "open," it will further promote the development of AI infrastructure.
Regarding whether the Scale-up Ethernet collaboration framework between OpenAI and other companies will lead to mandatory technology upgrade requirements for Taiwanese server and network hardware manufacturers, Broadcom stated that OpenAI is only one of the 12 founding companies of the Scale-up Ethernet collaboration. Other companies include Meta, Microsoft, Oracle, Cisco, Arista, Juniper, Broadcom, NVIDIA, AMD, and even more companies will join and provide feedback. Therefore, the specification setting will not be dominated by a single manufacturer, so the Taiwanese supply chain does not need to worry too much about response issues.
